Microsoft launches seven specialized AI models to undercut Claude and OpenAI costs
The new MAI model family offers high performance for specific tasks like coding and security at a fraction of the price of leading competitors.

The short version
- MAI Code 2 offers 96 percent of the performance of Anthropic's Claude Opus 5.5 at just 5 percent of the cost.
- MAI Cyber 2 Flash matches GPT-6 Astra's quality at less than 2 percent of the cost per task on specific benchmarks.
- Microsoft introduced the Maia 200 chip, which delivers up to 1.7 times more tokens per watt than leading GPU infrastructure.
- Decision 1.1 is a model routing tool that directs tasks to larger models only when necessary, making it 79 times cheaper than GPT-Sol 6.1.
- MAI Licensed Weights now allows businesses to customize models with proprietary data within their own Azure environments.
Why it matters
For businesses, this shift means they no longer have to pay premium prices for general-purpose AI when a smaller, specialized model can do the job. It enables the scaling of automated agents by significantly lowering the 'token cost'—the price of processing information—for high-volume tasks.
A shift toward specialized AI
Microsoft recently unveiled a new family of models called MAI, designed to compete directly with major players like OpenAI and Anthropic. Rather than relying on a single, massive system to handle every possible request, Microsoft is moving toward a strategy of specialization. This means different models are built to excel at specific tasks like coding, cybersecurity, image editing, and speech transcription.
The company announced these models during its Ignite conference on November 17. The lineup includes several distinct versions, such as MAI Code 2 for software development and MAI Cyber 2 Flash for security monitoring. This approach is intended to help enterprises manage the rising costs of running automated systems continuously. By using a model tailored for a specific job, companies can achieve better results while spending less money.
As businesses begin to deploy automated agents that perform repetitive tasks, the cost of 'tokens'—the basic units of text or data processed by AI—has become a significant hurdle. Microsoft is positioning these new models as a way to make these systems more affordable for large-scale use. The goal is to provide high-quality performance without the high price tag associated with the most powerful general-purpose models.
Cutting costs for developers
For software engineers and companies building applications, the new MAI Code 2 models are aimed at everyday enterprise coding tasks. Microsoft claims that MAI Code 2 delivers 98 percent of the coding performance of GPT-6.1 Sol but costs about 40 percent of the price. For even more intensive work, the company compares it to Anthropic’s Claude Opus 5.5, claiming it reaches 96 percent of that performance at only 5 percent of the cost.
There is also a lighter version called MAI Code 2 Flash for routine developer needs. This model is designed for simpler tasks like explaining existing code, making targeted edits, or fixing small bugs. It is already being integrated into GitHub Copilot and is available in a limited preview through Microsoft Foundry. This allows developers to choose the right level of power for the specific coding problem they are trying to solve.
By offering these different tiers, Microsoft is trying to give developers more flexibility in their workflows. A developer might use the Flash version for quick questions throughout the day and reserve the more powerful MAI Code 2 for complex architectural changes. This distinction helps teams optimize their budgets while maintaining high productivity. The strategy focuses on matching the quality of the best models in the market while drastically reducing the bill for the user.
Security and data privacy
The MAI Cyber 2 Flash model is specifically built for continuous cybersecurity monitoring and defensive operations. It is designed to identify and patch vulnerabilities in live production systems. Microsoft notes that this model builds on earlier systems that already account for almost 40 percent of scans in MDASH, which is their security assessment platform. Notably, 90 percent of scans that started with MAI continued using the system.
In terms of efficiency, Microsoft claims that MAI Cyber 2 Flash matches the quality of GPT-6 Astra while costing less than 2 percent of the price for each task on the CyberGymE2E benchmark. This highlights the company's focus on making high-end security features accessible to a wider range of organizations. By creating a model that specializes in defense, Microsoft allows security teams to run more frequent and comprehensive checks without worrying about prohibitive costs.
Beyond security, Microsoft is also introducing MAI Licensed Weights to address privacy and customization concerns. This feature allows businesses to take models and customize them using their own proprietary data. These customized models can then be run within their own private Azure environments. This gives companies more control over their data and allows them to create tools that are uniquely tuned to their specific business needs.
Hardware and routing efficiency
To support these models, Microsoft is highlighting its second-generation AI accelerator, the Maia 200 chip. This hardware is already in production and is used to run Microsoft's largest models. According to the company, the Maia 200 delivers up to 1.7 times more tokens per watt than the leading GPU infrastructure currently available in the market.
Improving the hardware is just one part of the efficiency puzzle; the company is also introducing software-based routing. The Decision 1.1 model acts as a router within applications to direct tasks to the appropriate AI. It sends simple tasks to smaller, faster models and only uses the larger, more expensive models when absolutely necessary. Microsoft states that Decision 1.1 is 50 times faster and 79 times cheaper than GPT-Sol 6.1.
This combination of specialized models, smarter routing, and custom hardware aims to give businesses more control over their AI spending. By optimizing how tasks are distributed and how the underlying hardware processes them, Microsoft hopes to lower the barrier to entry for advanced AI. This multi-layered approach is designed to make it easier for companies to scale their automated systems without seeing their costs spiral out of control as they grow.
Multimedia and transcription
The MAI family also extends into image and speech technologies to provide more diverse options for enterprise use. MAI Image 2.7 is designed for image editing and reportedly ranks first in that category on Artificial Analysis. Microsoft claims it performs at one-fifth of the price of OpenAI’s GPT Image 2.6 Sunburst, while also ranking fourth in general text-to-image generation performance.
For audio needs, MAI Realtime 1 is Microsoft’s first speech-to-speech model, which is built for natural conversations with low latency. It is priced starting at $0.01 per audio minute. For transcription, the Dragon Copilot Transcribe tool is based on MAI Transcribe 2. This tool currently powers 10 million patient encounters every month and has reportedly halved the error rates for multilingual transcription.
These diverse releases show that Microsoft is attempting to cover every major category of AI utility. Whether a company needs to transcribe medical records, edit marketing images, or build a low-latency voice assistant, Microsoft is providing a specific tool for that job. By offering these specialized models, they are moving away from a 'one size fits all' approach toward a more modular and cost-effective ecosystem for their enterprise customers.
What to watch
- How other companies like Anthropic and OpenAI respond to these aggressive price cuts.
- Whether the Maia 200 chip becomes the standard for running Microsoft's largest models.
- How many enterprises actually switch from general models to these specialized MAI models.
Try it
Check the Microsoft Foundry website to see if you can access the limited preview of MAI Code 2 Flash for your coding projects.
Sources and links
First reported by interestingengineering.com. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.